Please add basic semantic parsing features

[rikiyanai]

what would make this tool useful:

• Common over claims,
• Project based groupings,
• process failures
• local llm analyzing surface
• git history matching
• md file matching
• obsidian plugin

[daaain]

Hey @rikiyanai, can you please explain in a bit more detail what you have in mind?

Maybe take one or two use cases from your list and outline the process?

[cboos]

I put my psychic hat on, or rather asked Claude to wear it, and here's what he thinks this was about...

---

My decoding attempt:

Title "basic semantic parsing features" + checklist reads as a wish list for turning the transcript viewer from a log reader into a queryable knowledge base. Each bullet is a different way to extract and correlate meaning from transcripts.

My guesses, bullet by bullet:

Bullet	Reading
"Common over claims"	Probably "common overclaims" — a hallucination/fact-check detector. Find patterns where Claude repeatedly asserts something unsupported. Alt reading: "prefer common patterns over individual claims" (aggregate analytics).
"Project based groupings"	Cross-project semantic clustering — sessions that worked on the same feature even if they span different project dirs. We already group by ~/.claude/projects/*/, this wants something deeper.
"process failures"	Failure forensics view — detect sessions where Claude gave up, looped, or where the user had to intervene. "Show me the painful sessions."
"local llm analyzing surface"	Export/API surface for another local LLM to consume. A structured transcript dump suitable for feeding into a local model for downstream analysis. Pairs naturally with --detail low (or minimal) — which keeps assistant responses and key subagent signals like WebSearch summaries. user-only would starve the analyzer of context.
"git history matching"	Correlate session events with git commits. "Which session produced commit abc123? What was the prompt that led to this change?"
"md file matching"	Same idea but for markdown files — which session edited work/foo.md?
"obsidian plugin"	Integrate into Obsidian vault so transcripts become first-class notes that can be cross-referenced from other notes.

The coherent picture: the user wants transcripts to connect — to code (git), to docs (md), to their PKM (Obsidian), and to other LLMs (local LLM surface). Currently the HTML is a terminal view; they want it to be a hub.

The "semantic parsing" in the title then means: extract structured signal (what files touched, what commits produced, what claims made) from the raw chat prose, so downstream tools can query it.

---

Claude also noted that the right move was @daaain's asking for clarification ;-) It stands that the above ideas are interesting in their own right, and as Claude noted, in line with my own thinking (for the --detail low / --format md features).

[daaain]

Yeah, I think the best thing to do for this library is to be a solid building block, correctly parsing and semantically understanding the logs (as I think a lot of your recent work @cboos would fall under that, taking a lot of the new raw logs and giving them richer meaning / metadata!) and making them available in convenient formats for other tools to build on.

I personally don't have the bandwidth to increase the scope, even just keeping up with new features of Claude Code seem to be a decent amount of work. There's also space for the library to get better performance. I also considered things like some plugin system with hooks, a Python (and maybe other language) bindings so it could be used more as an SDK, etc but I think between the SQLite db and watching the file system we already have some pretty universal interfaces to build on. But of course happy to take feedback if there's a simple enough opportunity for us to improve this interoperability!

[cboos]

Bullet	Reading
"obsidian plugin"	Integrate into Obsidian vault so transcripts become first-class notes that can be cross-referenced from other notes.

I'm thinking of adding some additional flexibility to improve Obsidian support (or similar Markdown-based KM tools).

• With --output, the Markdown files are generated under a target directory, but under a "flat" structure, mirroring the one in ~/.claude/projects (already there)
• With the --expand-paths flag, a target project could be "projected back" onto its real path: -home-joe-project-A => <output>/home/joe/project/A.
  • note that this also goes into the direction of "Project based groupings": if related projects were grouped in the filesystem, they will now also be grouped in the index
• With --filter-path, only the matching projects would be selected, and then the corresponding path would be truncated from the output.

All the three together:

claude-code-log --output ~/Documents/Obsidian/ClaudeProjects --expand-paths --filter-path /home/joe --format md --detail low --compact

Would result in rendered transcript being generated into:

~/Documents/Obsidian/ClaudeProjects/project/A/<session>.low.md

[daaain]

Oh, that Obsidian output sounds great! Wonder if it would be worth also outputting index files with Obsidian style wiki links 🤔

[cboos]

Well, considering that Obsidian has no problem with standard Markdown links, I don't think we need to do anything special there.

[rikiyanai]

now with deep seek flash, I regularly mine transcripts to keep my failure log updated, I think its an opportunity for these kinds of out of scope features

[cboos]

Bullet	Reading
"git history matching"	Correlate session events with git commits. "Which session produced commit abc123? What was the prompt that led to this change?"

grep/rg <commit> <rendered-projects-dir> -t md could easily locate the session and the closest prompt.
#151 could make that easier, but even now with <rendered-projects-dir> == ~/.claude/projects, this would work.

However, we could still enrich the markup with links to the remote (branch -> remote, remote -> GitHub/GitLab detection on the URL). Local commits would be rendered as 7c2e6f6, those found on the remote as 7c2e6f6.